How much does DigitalOcean Cuttlefish cost?

DigitalOcean's Cuttlefish Terms say the app is provided at no charge, while Managed Agents and other services used through it are charged. The pricing page lists $0.044 per vCPU-hour for CPU actually used, $0.0095 per GB-hour for memory, $0.05 per GiB-month for storage and $0.01 per GiB for data sent out of the sandbox, and says there is no per-session or per-product spend limit. Monthly plans are listed at $50 and $200. We tested nothing and compute no bill.

Is Cuttlefish ready for production use?

DigitalOcean's Cuttlefish Terms say the app is in public preview and "not recommended for production or sensitive workloads", and recommend using test data. They say no service-level agreement applies to the app and that DigitalOcean may conclude the preview at any time without advance notice. For Harness Runtime and Action Gateway, the Managed Agents Public Preview Terms (updated November 11, 2026) say the preview is provided "entirely without service level agreements (SLAs), uptime commitments, or performance guarantees". The pages we read state no general-availability date.

What do I need to use Cuttlefish?

DigitalOcean's how-to lists an iPhone running iOS 26.5 or later and a DigitalOcean account on a team with access to Managed Agents. Sign-in asks for six permissions (OAuth access scopes) and fails if you clear any. The availability page marks one region for Harness Runtime, RIC1, and does not say whether others are planned. The pages we read describe an iPhone app.

Does Cuttlefish ask for approval on every agent action?

No. DigitalOcean's approvals page says an action is gated only if the environment's permission policy says so, and that with a default of allow and no ask rules a session "never raises an approval". The how-to says Cuttlefish shows a request when the policy is set to ask. If the permissions block is left out, the permissions page says the agent keeps a built-in default that "varies by adapter". The how-to adds that you cannot answer a question the agent asks inside the app.

Is there a spending limit on Harness Runtime?

The pricing page says: "There is no per-session or per-product spend limit for Harness Runtime." Per the pricing and limits pages, pausing a session stops compute charges, retained checkpoints keep accruing storage charges, and sessions pause when all your balances are empty. The Public Preview Terms add that retained storage keeps accruing charges while the account is blocked. CPU is billed on use and falls to zero while an agent waits; memory is billed on what is allocated.

What dates do the pages give for Cuttlefish, and what stage is it in?

DigitalOcean's blog and release notes are dated November 5, 2026, and Apple's App Store data shows versions from October 28 to November 9. DigitalOcean calls the app a public preview. Its Terms also use the phrase "a Beta Service for purposes of the Agreement", a label tied to a main service agreement that we did not read. The pages we read state no general-availability date.

DigitalOcean Cuttlefish: Price, Limits, Preview

The short version

DigitalOcean's Cuttlefish is an iPhone app for allowing or denying what an AI coding agent does. DigitalOcean calls it a public preview, "not recommended for production or sensitive workloads", with no service-level agreement (an uptime promise) for the app or the service behind it. The Terms say the app is provided at no charge; usage is billed, and the pricing page says there is "no per-session or per-product spend limit".

In plain words

Think of Cuttlefish as a remote control for a job that runs on someone else's computer (our comparison, not DigitalOcean's). The job is an AI coding agent, a program that writes and changes software. It runs in a sandbox, a virtual computer on DigitalOcean's cloud that DigitalOcean calls "hardware-isolated". When the agent needs a yes or no, such as before running a command, the app shows the exact step on your iPhone.

DigitalOcean prices CPU, memory, storage and outbound data separately: CPU follows use, memory is billed on what is allocated, storage on what is held, and saved copies of a session keep billing while it is paused. Its pages call the whole thing a public preview. It matters if your team already uses DigitalOcean for this. We tested nothing.

DigitalOcean, a cloud company, has an iPhone app called Cuttlefish. The agent runs on DigitalOcean's cloud, so it can keep working while you are away from your desk. When it wants to do something, such as run a command, the app lets you tap allow or deny. DigitalOcean's Cuttlefish Terms call the app a public preview, "not recommended for production or sensitive workloads", and say no service-level agreement (a written uptime promise) applies to it.

The audience is narrow. The how-to page asks for "An iPhone running iOS 26.5 or later" and "A DigitalOcean account on a team with access to Managed Agents." Managed Agents is DigitalOcean's name for the service that runs the agents; Harness Runtime, the part that runs the sandboxes, is what the prices and limits below belong to. If your team does not run coding agents there, this is mainly a look at what a phone approval step involves.

What does the app do?

A coding agent can keep working while you are away (the blog says agents run for "minutes, hours, or overnight"), and it stops to ask before certain steps, such as running a command. Cuttlefish shows that request with "the full command, file path, or diff", and you tap allow or deny. You can also send a correction, pause or stop the run, and browse workspace files.

The work stays on DigitalOcean's side, in a sandbox (a virtual computer; DigitalOcean's pages call these "hardware-isolated microVM sandboxes"): the how-to says "sessions keep running when your phone is locked or offline".

On dates: DigitalOcean's blog is printed "Updated: November 5, 2026", and its release notes list the app under 5 November. A DigitalOcean press release dated November 5 says the app "is available today on the App Store for Managed Agents customers". Apple's App Store data lists versions from October 28 to a version 0.1.2 dated November 9. The pages we read do not say when the preview began or when it ends.

When does the agent actually ask you?

Only when the environment's permission policy says so. A permission policy is a list of rules: allow (the agent acts alone), ask (it waits for a person) or deny (blocked). DigitalOcean's approvals page says that with a default of allow and no ask rules, a session "never raises an approval". If the default is left out, the permissions page says it "resolves to ask". If the whole block is left out, the agent "keeps whatever built-in default it ships with, which varies by adapter" (an adapter connects each coding agent to the platform). In our reading, the app is only as useful as the policy someone wrote.

The Cuttlefish Terms add a caution. They call approval prompts, risk labels and device authentication "conveniences" that "may not identify every risky action". The how-to (verified 5 Nov 2026, four days before the app's version 0.1.2 is dated) says allowing a request the app marks as risky needs Face ID or the device passcode, and that if the device has neither set up, "Cuttlefish asks you to confirm instead." Apple's notes for version 0.1.2 (dated November 9 in Apple's data) say: "Choose how often Face ID asks when you allow a request", or turn it off. The pages we read do not say whether that setting covers requests the app marks as risky.

Two more limits, from the how-to. It says you "can't answer" a question the agent asks inside the app; you answer it from doctl (DigitalOcean's command-line tool) or dismiss it. In the background, notifications "can arrive many minutes late", and none arrive after you force-quit the app. The approvals page calls an agent waiting on an approval that nobody answers "a stalled session".

What does it cost?

The Cuttlefish Terms (the app's Terms) say "The App is provided at no charge". Managed Agents and other services used through it are charged: the Managed Agents Public Preview Terms (the separate Terms for the service behind the app) add "The Preview Offering is a paid service." Every price below is printed on DigitalOcean's pricing page and can change.

Resource	Price as printed	Billing basis (DigitalOcean's wording)
CPU (a vCPU is one virtual processor core)	$0.044 per vCPU-hour	Actual CPU consumed
Memory	$0.0095 per GB-hour	Actual memory allocated over time
Session storage (a GiB is about a gigabyte)	$0.05 per GiB-month	Actual capacity held over time
Data sent out to the internet (egress)	$0.01 per GiB	Outbound traffic from the sandbox to the public internet

Of the compute parts, only CPU follows consumption: while an agent waits and uses no CPU, "its CPU charge falls to zero". Memory is billed on what is allocated, and storage on what is held. "Pausing a session stops compute charges", but "Retained checkpoints" (saved copies of the session) "continue to accrue storage charges while a session is paused".

The page prints a size table, priced "at full allocation", from XSmall (1 vCPU, 1 GB) at $0.0535 per hour to XLarge (16 vCPUs, 32 GB) at $1.008 per hour. Our check of the default Medium size (2 vCPUs, 4 GB): 2 x $0.044 + 4 x $0.0095 = $0.126 per hour, which matches the printed row. We derive no bill from it; storage, data sent out, tools and models are priced separately.

In our reading, the key budget sentence is: "There is no per-session or per-product spend limit for Harness Runtime." The page also warns that an agent stuck in a loop, using tools that require prepayment, can use up your balances quickly (our paraphrase of its warning). The Public Preview Terms (section 3.5) say that when the balance reaches zero DigitalOcean will "pause running MicroVMs and take a snapshot", and that retained data "continues to accrue storage charges for as long as it is retained".

Monthly plans, in public preview, are listed at $50 (Pro, 15% discount) and $200 (Team, 20%). The page says the fee funds a subscription balance that usage draws from, and that after it is used up usage "continues at standard rates" from other balances.

What are the limits?

Item	What DigitalOcean's pages say
Sessions at once	"up to 100 sessions at once per team, depending on your resource tier"; limits "depend on your account history and spend"
Active compute per session	"a monthly active compute cap of 744 hours" (our arithmetic: 31 days x 24 hours)
Idle sessions	"Idle suspend defaults to 15 minutes."
Where it runs	The availability table marks one region for Harness Runtime, RIC1. The page does not say whether others are planned.
Outbound network from the sandbox	"Sandbox egress is unrestricted unless the environment spec sets an allowlist." (our gloss: an allowlist names the hosts a session may reach)

What does "public preview" mean in the fine print?

The Cuttlefish Terms say DigitalOcean "may modify, suspend, or discontinue any feature of the App" at any time without notice, and may "conclude the Public Preview at any time without advance notice". They state "No Technical Support commitments and no Service Level Agreements (SLAs) apply to the App." That sentence is about the app. For the service behind it, DigitalOcean's Managed Agents Public Preview Terms (updated November 11, 2026) cover Harness Runtime and Action Gateway (DigitalOcean's catalog of tools an agent can call) together as the "Preview Offering". Section 2.2 says it is provided "entirely without service level agreements (SLAs), uptime commitments, or performance guarantees". The Cuttlefish Terms also call the app "a Beta Service for purposes of the Agreement", a label from DigitalOcean's main service agreement, which we did not read; we use the product label, public preview.

What happens to the data you send?

The Cuttlefish Privacy Policy says DigitalOcean collects your prompts and the files and photos you choose to upload, and that your content is sent only to the AI model set for your environment; depending on the model, DigitalOcean or the model's provider processes it. The policy adds that agents "may send content to tools you or your team connect", such as Slack or Jira. The Public Preview Terms say DigitalOcean "has enabled the DigitalOcean Insights feature for your account by default" (Insights is its monitoring feature), and the Harness Runtime data privacy page says Insights telemetry "may contain personal data" and is stored in a United States region.

The Harness Runtime data privacy page says content is kept "for as long as needed" under a retention schedule, with no number of days. It says DigitalOcean does not use your content to train any generalized AI or machine-learning model. For a third-party model, the page says that provider's own data-use and training practices apply. It also says data in a preview "may be lost" and that "you are responsible for maintaining your own backups". Section 2.4 of the Public Preview Terms says "DigitalOcean provides no guarantees regarding data durability." The data privacy page tells users not to submit health information covered by HIPAA (the US health-privacy law) or financial data covered by PCI-DSS (the payment-card data standard).

Where do the pages differ or stay silent?

Tool count. The blog says "20,000+ tools" for Managed Agents. The docs landing page and the Action Gateway page say Action Gateway gives access to "16,000+ tools". The pages we read do not say whether the two numbers count the same things.
Approvals for those tools. The approvals page (verified 22 Oct 2026) says a rule on one of those tools, if set to ask, "refuses the call with a message rather than prompting you". The older permissions page (verified 23 Sep 2026) says that for two of the agents (the claude-code and codex adapters) the gateway "pauses the tool call" and an approval request appears, and tells readers to "Verify approval behavior with the adapter you deploy". Neither says how the iPhone app handles it.
Balances. The pricing page lists an Agent Droplets plan, the "Inference and Agents balance" and the "account prepayment balance", and gives an order only for usage after a plan's balance is used up (promotional credits, then those two balances). The Public Preview Terms speak of a "single, account-level DigitalOcean platform wallet"; the pages we read do not say how the balances relate to it.
Not stated. The pages we read give no benchmark, success rate or time saved for supervising agents from a phone.

What to do this week

Four checks for an owner or team lead.

If your team does not use DigitalOcean Managed Agents, you can stop here: the how-to says the app needs an account on a team with access to it.
If it does, ask whoever set up your environments what the permission policy says for the default (the approvals page says a default of allow with no ask rules means a session never raises an approval). Name one person who answers approval requests, and decide what happens when nobody does.
Name one person who watches the prepaid balance, because the pricing page says there is no per-session or per-product spend limit.
In our reading, keep the work to test data, as the Terms recommend. Ask DigitalOcean in writing how many days session data is kept; the pages we read do not say.

Book an AI discovery call

How we know

On November 11, 2026 we read the sixteen pages listed under Sources, with the date each prints where it prints one. We did not read DigitalOcean's Terms of Service Agreement, its Agent and Inference Balance Terms, or the two pages on paying that the pricing page links. We did not open the app or test anything. Pages can change.

DigitalOcean's Cuttlefish iPhone App for Managed Agents (Public Preview): What It Does, What It Costs, and What the Docs Do Not Say